scikit-learn-intelex 2025.4.0__py313-none-manylinux_2_28_x86_64.whl

sklearnex/utils/_array_api.py

@@ -0,0 +1,82 @@
+# ==============================================================================
+# Copyright 2024 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+"""Tools to support array_api."""
+
+import numpy as np
+
+from daal4py.sklearn._utils import sklearn_check_version
+from onedal.utils._array_api import _get_sycl_namespace
+
+if sklearn_check_version("1.2"):
+    from sklearn.utils._array_api import get_namespace as sklearn_get_namespace
+
+
+def get_namespace(*arrays):
+    """Get namespace of arrays.
+
+    Introspect `arrays` arguments and return their common Array API
+    compatible namespace object, if any. NumPy 1.22 and later can
+    construct such containers using the `numpy.array_api` namespace
+    for instance.
+
+    This function will return the namespace of SYCL-related arrays
+    which define the __sycl_usm_array_interface__ attribute
+    regardless of array_api support, the configuration of
+    array_api_dispatch, or scikit-learn version.
+
+    See: https://numpy.org/neps/nep-0047-array-api-standard.html
+
+    If `arrays` are regular numpy arrays, an instance of the
+    `_NumPyApiWrapper` compatibility wrapper is returned instead.
+
+    Namespace support is not enabled by default. To enabled it
+    call:
+
+      sklearn.set_config(array_api_dispatch=True)
+
+    or:
+
+      with sklearn.config_context(array_api_dispatch=True):
+          # your code here
+
+    Otherwise an instance of the `_NumPyApiWrapper`
+    compatibility wrapper is always returned irrespective of
+    the fact that arrays implement the `__array_namespace__`
+    protocol or not.
+
+    Parameters
+    ----------
+    *arrays : array objects
+        Array objects.
+
+    Returns
+    -------
+    namespace : module
+        Namespace shared by array objects.
+
+    is_array_api : bool
+        True of the arrays are containers that implement the Array API spec.
+    """
+
+    sycl_type, xp, is_array_api_compliant = _get_sycl_namespace(*arrays)
+
+    if sycl_type:
+        return xp, is_array_api_compliant
+    elif sklearn_check_version("1.2"):
+        return sklearn_get_namespace(*arrays)
+    else:
+        return np, False

sklearnex/utils/parallel.py

@@ -0,0 +1,59 @@
+# ===============================================================================
+# Copyright 2023 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ===============================================================================
+
+import warnings
+from functools import update_wrapper
+
+from .._config import config_context, get_config
+
+
+class _FuncWrapper:
+    """Load the global configuration before calling the function."""
+
+    def __init__(self, function):
+        self.function = function
+        update_wrapper(self, self.function)
+
+    def with_config(self, config):
+        self.config = config
+        return self
+
+    def __call__(self, *args, **kwargs):
+        config = getattr(self, "config", None)
+        if config is None:
+            warnings.warn(
+                "`sklearn.utils.parallel.delayed` should be used with "
+                "`sklearn.utils.parallel.Parallel` to make it possible to propagate "
+                "the scikit-learn configuration of the current thread to the "
+                "joblib workers.",
+                UserWarning,
+            )
+            config = {}
+        with config_context(**config):
+            return self.function(*args, **kwargs)
+
+
+class _FuncWrapperOld:
+    """Load the global configuration before calling the function."""
+
+    def __init__(self, function):
+        self.function = function
+        self.config = get_config()
+        update_wrapper(self, self.function)
+
+    def __call__(self, *args, **kwargs):
+        with config_context(**self.config):
+            return self.function(*args, **kwargs)

sklearnex/utils/tests/test_validation.py

@@ -0,0 +1,238 @@
+# ==============================================================================
+# Copyright contributors to the oneDAL project
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+
+import numpy as np
+import numpy.random as rand
+import pytest
+
+from daal4py.sklearn._utils import sklearn_check_version
+from onedal.tests.utils._dataframes_support import (
+    _convert_to_dataframe,
+    get_dataframes_and_queues,
+)
+from sklearnex import config_context
+from sklearnex.tests.utils import DummyEstimator, gen_dataset
+from sklearnex.utils.validation import _check_sample_weight, validate_data
+
+# array_api support starts in sklearn 1.2, and array_api_strict conformance starts in sklearn 1.3
+_dataframes_supported = (
+    "numpy,pandas"
+    + (",dpctl" if sklearn_check_version("1.2") else "")
+    + (",array_api" if sklearn_check_version("1.3") else "")
+)
+
+
+@pytest.mark.parametrize("dtype", [np.float32, np.float64])
+@pytest.mark.parametrize(
+    "shape",
+    [
+        [16, 2048],
+        [2**16 + 3],
+        [1000, 1000],
+    ],
+)
+@pytest.mark.parametrize("ensure_all_finite", ["allow-nan", True])
+def test_sum_infinite_actually_finite(dtype, shape, ensure_all_finite):
+    est = DummyEstimator()
+    X = np.empty(shape, dtype=dtype)
+    X.fill(np.finfo(dtype).max)
+    X = np.atleast_2d(X)
+    X_array = validate_data(est, X, ensure_all_finite=ensure_all_finite)
+    assert type(X_array) == type(X)
+
+
+@pytest.mark.parametrize("dtype", [np.float32, np.float64])
+@pytest.mark.parametrize(
+    "shape",
+    [
+        [16, 2048],
+        [2**16 + 3],
+        [1000, 1000],
+    ],
+)
+@pytest.mark.parametrize("ensure_all_finite", ["allow-nan", True])
+@pytest.mark.parametrize("check", ["inf", "NaN", None])
+@pytest.mark.parametrize("seed", [0, 123456])
+@pytest.mark.parametrize(
+    "dataframe, queue",
+    get_dataframes_and_queues(_dataframes_supported),
+)
+def test_validate_data_random_location(
+    dataframe, queue, dtype, shape, ensure_all_finite, check, seed
+):
+    est = DummyEstimator()
+    rand.seed(seed)
+    X = rand.uniform(high=np.finfo(dtype).max, size=shape).astype(dtype)
+
+    if check:
+        loc = rand.randint(0, X.size - 1)
+        X.reshape((-1,))[loc] = float(check)
+
+    # column heavy pandas inputs are very slow in sklearn's check_array even without
+    # the finite check, just transpose inputs to guarantee fast processing in tests
+    X = _convert_to_dataframe(
+        np.atleast_2d(X).T,
+        target_df=dataframe,
+        sycl_queue=queue,
+    )
+
+    dispatch = {}
+    if sklearn_check_version("1.2") and dataframe != "pandas":
+        dispatch["array_api_dispatch"] = True
+
+    with config_context(**dispatch):
+
+        allow_nan = ensure_all_finite == "allow-nan"
+        if check is None or (allow_nan and check == "NaN"):
+            validate_data(est, X, ensure_all_finite=ensure_all_finite)
+        else:
+            type_err = "infinity" if allow_nan else "[NaN|infinity]"
+            msg_err = f"Input X contains {type_err}"
+            with pytest.raises(ValueError, match=msg_err):
+                validate_data(est, X, ensure_all_finite=ensure_all_finite)
+
+
+@pytest.mark.parametrize("dtype", [np.float32, np.float64])
+@pytest.mark.parametrize("ensure_all_finite", ["allow-nan", True])
+@pytest.mark.parametrize("check", ["inf", "NaN", None])
+@pytest.mark.parametrize("seed", [0, 123456])
+@pytest.mark.parametrize(
+    "dataframe, queue",
+    get_dataframes_and_queues(_dataframes_supported),
+)
+def test_validate_data_random_shape_and_location(
+    dataframe, queue, dtype, ensure_all_finite, check, seed
+):
+    est = DummyEstimator()
+    lb, ub = 32768, 1048576  # lb is a patching condition, ub 2^20
+    rand.seed(seed)
+    X = rand.uniform(high=np.finfo(dtype).max, size=rand.randint(lb, ub)).astype(dtype)
+
+    if check:
+        loc = rand.randint(0, X.size - 1)
+        X[loc] = float(check)
+
+    X = _convert_to_dataframe(
+        np.atleast_2d(X).T,
+        target_df=dataframe,
+        sycl_queue=queue,
+    )
+
+    dispatch = {}
+    if sklearn_check_version("1.2") and dataframe != "pandas":
+        dispatch["array_api_dispatch"] = True
+
+    with config_context(**dispatch):
+
+        allow_nan = ensure_all_finite == "allow-nan"
+        if check is None or (allow_nan and check == "NaN"):
+            validate_data(est, X, ensure_all_finite=ensure_all_finite)
+        else:
+            type_err = "infinity" if allow_nan else "[NaN|infinity]"
+            msg_err = f"Input X contains {type_err}."
+            with pytest.raises(ValueError, match=msg_err):
+                validate_data(est, X, ensure_all_finite=ensure_all_finite)
+
+
+@pytest.mark.parametrize("dtype", [np.float32, np.float64])
+@pytest.mark.parametrize("check", ["inf", "NaN", None])
+@pytest.mark.parametrize("seed", [0, 123456])
+@pytest.mark.parametrize(
+    "dataframe, queue",
+    get_dataframes_and_queues(_dataframes_supported),
+)
+def test__check_sample_weight_random_shape_and_location(
+    dataframe, queue, dtype, check, seed
+):
+    # This testing assumes that array api inputs to validate_data will only occur
+    # with sklearn array_api support which began in sklearn 1.2. This would assume
+    # that somewhere upstream of the validate_data call, a data conversion of dpnp,
+    # dpctl, or array_api inputs to numpy inputs would have occurred.
+
+    lb, ub = 32768, 1048576  # lb is a patching condition, ub 2^20
+    rand.seed(seed)
+    shape = (rand.randint(lb, ub), 2)
+    X = rand.uniform(high=np.finfo(dtype).max, size=shape).astype(dtype)
+    sample_weight = rand.uniform(high=np.finfo(dtype).max, size=shape[0]).astype(dtype)
+
+    if check:
+        loc = rand.randint(0, shape[0] - 1)
+        sample_weight[loc] = float(check)
+
+    X = _convert_to_dataframe(
+        X,
+        target_df=dataframe,
+        sycl_queue=queue,
+    )
+    sample_weight = _convert_to_dataframe(
+        sample_weight,
+        target_df=dataframe,
+        sycl_queue=queue,
+    )
+
+    dispatch = {}
+    if sklearn_check_version("1.2") and dataframe != "pandas":
+        dispatch["array_api_dispatch"] = True
+
+    with config_context(**dispatch):
+
+        if check is None:
+            X_out = _check_sample_weight(sample_weight, X)
+            if dispatch:
+                assert type(X_out) == type(X)
+            else:
+                assert isinstance(X_out, np.ndarray)
+        else:
+            msg_err = "Input sample_weight contains [NaN|infinity]"
+            with pytest.raises(ValueError, match=msg_err):
+                X_out = _check_sample_weight(sample_weight, X)
+
+
+@pytest.mark.parametrize("dtype", [np.float32, np.float64])
+@pytest.mark.parametrize(
+    "dataframe, queue",
+    get_dataframes_and_queues(_dataframes_supported),
+)
+def test_validate_data_output(dtype, dataframe, queue):
+    # This testing assumes that array api inputs to validate_data will only occur
+    # with sklearn array_api support which began in sklearn 1.2. This would assume
+    # that somewhere upstream of the validate_data call, a data conversion of dpnp,
+    # dpctl, or array_api inputs to numpy inputs would have occurred.
+    est = DummyEstimator()
+    X, y = gen_dataset(est, queue=queue, target_df=dataframe, dtype=dtype)[0]
+
+    dispatch = {}
+    if sklearn_check_version("1.2") and dataframe != "pandas":
+        dispatch["array_api_dispatch"] = True
+
+    with config_context(**dispatch):
+        X_out, y_out = validate_data(est, X, y)
+        # check sklearn validate_data operations work underneath
+        X_array = validate_data(est, X, reset=False)
+
+    for orig, first, second in ((X, X_out, X_array), (y, y_out, None)):
+        if dispatch:
+            assert type(orig) == type(
+                first
+            ), f"validate_data converted {type(orig)} to {type(first)}"
+            if second is not None:
+                assert type(orig) == type(
+                    second
+                ), f"from_array converted {type(orig)} to {type(second)}"
+        else:
+            # array_api_strict from sklearn < 1.2 and pandas will convert to numpy arrays
+            assert isinstance(first, np.ndarray)
+            assert second is None or isinstance(second, np.ndarray)

sklearnex/utils/validation.py

@@ -0,0 +1,208 @@
+# ===============================================================================
+# Copyright 2022 Intel Corporation
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ===============================================================================
+
+import numbers
+
+import scipy.sparse as sp
+from sklearn.utils.validation import _assert_all_finite as _sklearn_assert_all_finite
+from sklearn.utils.validation import _num_samples, check_array, check_non_negative
+
+from daal4py.sklearn._utils import daal_check_version, sklearn_check_version
+
+from ._array_api import get_namespace
+
+if sklearn_check_version("1.6"):
+    from sklearn.utils.validation import validate_data as _sklearn_validate_data
+
+    _finite_keyword = "ensure_all_finite"
+
+else:
+    from sklearn.base import BaseEstimator
+
+    _sklearn_validate_data = BaseEstimator._validate_data
+    _finite_keyword = "force_all_finite"
+
+
+if daal_check_version((2024, "P", 700)):
+    from onedal.utils.validation import _assert_all_finite as _onedal_assert_all_finite
+
+    def _onedal_supported_format(X, xp):
+        # array_api does not have a `strides` or `flags` attribute for testing memory
+        # order. When dlpack support is brought in for oneDAL, general support for
+        # array_api can be enabled and the hasattr check can be removed.
+        # _onedal_supported_format is therefore conservative in verifying attributes and
+        # does not support array_api. This will block onedal_assert_all_finite from being
+        # used for array_api inputs but will allow dpnp ndarrays and dpctl tensors.
+        # only check contiguous arrays to prevent unnecessary copying of data, even if
+        # non-contiguous arrays can now be converted to oneDAL tables.
+        return (
+            X.dtype in [xp.float32, xp.float64]
+            and hasattr(X, "flags")
+            and (X.flags["C_CONTIGUOUS"] or X.flags["F_CONTIGUOUS"])
+        )
+
+else:
+    from daal4py.utils.validation import _assert_all_finite as _onedal_assert_all_finite
+    from onedal.utils._array_api import _is_numpy_namespace
+
+    def _onedal_supported_format(X, xp):
+        # daal4py _assert_all_finite only supports numpy namespaces, use internally-
+        # defined check to validate inputs, otherwise offload to sklearn
+        return X.dtype in [xp.float32, xp.float64] and _is_numpy_namespace(xp)
+
+
+def _sklearnex_assert_all_finite(
+    X,
+    *,
+    allow_nan=False,
+    input_name="",
+):
+    # size check is an initial match to daal4py for performance reasons, can be
+    # optimized later
+    xp, _ = get_namespace(X)
+    if X.size < 32768 or not _onedal_supported_format(X, xp):
+        if sklearn_check_version("1.1"):
+            _sklearn_assert_all_finite(X, allow_nan=allow_nan, input_name=input_name)
+        else:
+            _sklearn_assert_all_finite(X, allow_nan=allow_nan)
+    else:
+        _onedal_assert_all_finite(X, allow_nan=allow_nan, input_name=input_name)
+
+
+def assert_all_finite(
+    X,
+    *,
+    allow_nan=False,
+    input_name="",
+):
+    _sklearnex_assert_all_finite(
+        X.data if sp.issparse(X) else X,
+        allow_nan=allow_nan,
+        input_name=input_name,
+    )
+
+
+def validate_data(
+    _estimator,
+    /,
+    X="no_validation",
+    y="no_validation",
+    **kwargs,
+):
+    # force finite check to not occur in sklearn, default is True
+    # `ensure_all_finite` is the most up-to-date keyword name in sklearn
+    # _finite_keyword provides backward compatability for `force_all_finite`
+    ensure_all_finite = kwargs.pop("ensure_all_finite", True)
+    kwargs[_finite_keyword] = False
+
+    out = _sklearn_validate_data(
+        _estimator,
+        X=X,
+        y=y,
+        **kwargs,
+    )
+
+    check_x = not isinstance(X, str) or X != "no_validation"
+    check_y = not (y is None or isinstance(y, str) and y == "no_validation")
+
+    if ensure_all_finite:
+        # run local finite check
+        allow_nan = ensure_all_finite == "allow-nan"
+        # the return object from validate_data can be a single
+
+        # element (either x or y) or both (as a tuple). An iterator along with
+        # check_x and check_y can go through the output properly without
+        # stacking layers of if statements to make sure the proper input_name
+        # is used
+        arg = iter(out if isinstance(out, tuple) else (out,))
+        if check_x:
+            assert_all_finite(next(arg), allow_nan=allow_nan, input_name="X")
+        if check_y:
+            assert_all_finite(next(arg), allow_nan=allow_nan, input_name="y")
+
+    if check_y and "dtype" in kwargs:
+        # validate_data does not do full dtype conversions, as it uses check_X_y
+        # oneDAL can make tables from [int32, float32, float64], requiring
+        # a dtype check and conversion. This will query the array_namespace and
+        # convert y as necessary. This is important especially for regressors.
+        dtype = kwargs["dtype"]
+        if not isinstance(dtype, (tuple, list)):
+            dtype = tuple(dtype)
+
+        outx, outy = out if check_x else (None, out)
+        if outy.dtype not in dtype:
+            yp, _ = get_namespace(outy)
+            # use asarray rather than astype because of numpy support
+            outy = yp.asarray(outy, dtype=dtype[0])
+            out = (outx, outy) if check_x else outy
+
+    return out
+
+
+def _check_sample_weight(
+    sample_weight, X, dtype=None, copy=False, ensure_non_negative=False
+):
+
+    n_samples = _num_samples(X)
+    xp, _ = get_namespace(X)
+
+    if dtype is not None and dtype not in [xp.float32, xp.float64]:
+        dtype = xp.float64
+
+    if sample_weight is None:
+        if hasattr(X, "device"):
+            sample_weight = xp.ones(n_samples, dtype=dtype, device=X.device)
+        else:
+            sample_weight = xp.ones(n_samples, dtype=dtype)
+    elif isinstance(sample_weight, numbers.Number):
+        if hasattr(X, "device"):
+            sample_weight = xp.full(
+                n_samples, sample_weight, dtype=dtype, device=X.device
+            )
+        else:
+            sample_weight = xp.full(n_samples, sample_weight, dtype=dtype)
+    else:
+        if dtype is None:
+            dtype = [xp.float64, xp.float32]
+
+        params = {
+            "accept_sparse": False,
+            "ensure_2d": False,
+            "dtype": dtype,
+            "order": "C",
+            "copy": copy,
+            _finite_keyword: False,
+        }
+        if sklearn_check_version("1.1"):
+            params["input_name"] = "sample_weight"
+
+        sample_weight = check_array(sample_weight, **params)
+        assert_all_finite(sample_weight, input_name="sample_weight")
+
+        if sample_weight.ndim != 1:
+            raise ValueError("Sample weights must be 1D array or scalar")
+
+        if sample_weight.shape != (n_samples,):
+            raise ValueError(
+                "sample_weight.shape == {}, expected {}!".format(
+                    sample_weight.shape, (n_samples,)
+                )
+            )
+
+    if ensure_non_negative:
+        check_non_negative(sample_weight, "`sample_weight`")
+
+    return sample_weight